Fully automated deep learning system for osteoporosis screening using chest computed tomography images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Osteoporosis affects 200 million people globally, stemming from bone metabolism irregularities.
• Early detection of osteoporosis is crucial for public health management.
• Deep learning (DL) shows promise in medical imaging for bone mineral density (BMD) assessment.

Purpose of the Study:

• To propose an automated DL framework for BMD assessment.
• The framework integrates localization, segmentation, and ternary classification.
• Utilize dominant convolutional neural networks (CNNs) for BMD assessment.

Main Methods:

• A retrospective study of 2,274 patients undergoing chest CT scans.
• Developed an automated segmentation model using VB-Net networks.
• Employed DenseNet, ResNet-18, and ResNet-50 for BMD classification, evaluated using ROC curves and AUC.

Main Results:

• Segmentation model achieved a Dice similarity coefficient > 0.93.
• DenseNet and ResNet-18 showed excellent diagnostic performance for osteoporosis and osteopenia (AUCs up to 0.96).
• ResNet-50 had suboptimal performance for osteopenia (AUC 0.76); DenseNet performance was more sensitive to tube voltage variations.

Conclusions:

• The developed DL framework provides an effective and efficient method for opportunistic osteoporosis screening.
• Utilizes existing chest CT scans, avoiding additional costs and radiation exposure.
• Offers a valuable tool for early detection and management of osteoporosis.